Flow¶
Flow-based generative model
NICE¶
NICE: Non-linear Independent Components Estimation (2014)
workshop paper at ICLR 2015
optimize distribution by integration
General coupling layer
merits:
- Exact latent-variable inference and log-likelihood evaluation
- Efficient inference and efficient synthesis
- Useful latent space for downstream tasks 细水长flow之NICE：流模型的基本概念与实现
realNVP¶
real-valued non-volume preserving
generalize coupling layer, add convolution
Glow¶
Glow: Generative Flow with Invertible 1×1 Convolutions (2018)
extends from NICE and RealNVP
addition of a reversible 1x1 convolution, as well as removing other components, simplifying the architecture overall
Glow: Better Reversible Generative Models - OpenAI
BIG disadvantage: computation cost for training is too high
256x256 high resolution face is in trained with 40 GPU for about a week. ~1 year for 1 GPU (:3 J L)
github issue #37: anyone reproduced the celeba-HQ results in the paper